| The success of Generative Adversarial Networks(GAN)is largely due to the abundance of raw data resources available online.Like most neural networks,GANs have always followed a basic principle: the more training data there is,the better the model’s performance will be.This is because each independent GAN is composed of two mutually learning neural networks:a generator that generates data,and a discriminator that strives to distinguish between generated and real data.During training,the discriminator guides the generator by providing frame-byframe feedback through backpropagation to help improve the realism of the synthesized images.In style transfer applications,examples are often scarce and difficult to obtain,while collecting the dataset itself is a time-consuming and laborious task.Taking Cycle-Consistent Adversarial Networks(Cycle GAN)as an example,it can perform image transformations between two different data domains without requiring one-to-one annotated data.However,due to the presence of cycle loss,training a Cycle GAN model requires a considerable amount of data.To enable the Cycle GAN model to be trained with limited data,this paper proposes the following approach:First,this paper analyzes the reason why training the original GAN model with a small amount of data leads to overfitting,and then analyzes the feasibility of training GAN models with limited data.GAN models use the concept of adversarial learning to train the generator and discriminator alternately.The features that the generator learns in adversarial learning depend on the feedback from the discriminator.In the case of a small dataset,the discriminator will overfit,causing the generator to fail to obtain effective feedback from the discriminator.Therefore,to train GAN models with limited data,the prediction values of the discriminator can be constrained to prevent them from diverging too much.This also leads to the regularization of the discriminator in this paper.For the Cycle GAN model for image transformation tasks,this paper proposes two regularization methods: L1 regularization based on the original GAN loss function and L2 regularization based on the least squares loss.The above two regularization methods are analyzed separately.The regularization method based on the original GAN model restricts the prediction values of the discriminator to around 0.5.The regularization based on the least squares loss restricts the prediction values of the discriminator to around the median of the prediction domain.Through mathematical derivation,it can be shown that the loss function of this regularization model is related to the Le Cam distance and its related distance metrics.Minimizing the loss function is equivalent to minimizing the Le Cam distance between the generated data distribution and the real data distribution.Comparative experiments are conducted between the Le Cam distance-based Cycle GAN and the original Cycle GAN model,and the dataset is divided into different proportions: 100%,50%,and 20% of the original dataset.The experimental results show that the original Cycle GAN model overfits the limited training data,causing the predicted results of real and generated images to diverge rapidly.After regularization is introduced,the difference between the predicted results of real and generated images is penalized to keep the predicted results within a specific range.The experimental results verify that the regularization model introduced in this paper can help the predicted results of the discriminator converge to a stable point more quickly.The regularization model proposed in this paper performs better on all three datasets,and the FID score is lower compared to the original Cycle GAN model. |
